Data Scaling Laws in Imitation Learning (Robotic Manipulation)

Research into how, in robot imitation learning, generalization improves as training data grows.

This is a paper from Tsinghua University's Yang Gao group (with partners including the Shanghai Qi Zhi Institute and the Shanghai AI Laboratory), released in October 2024 and selected as an oral presentation at ICLR 2025. Using the handheld capture device UMI (Universal Manipulation Interface), the team collected more than 40,000 demonstrations and ran more than 15,000 real-robot tests to study how a single-task policy's generalization to new environments and new objects scales with data. The finding: generalization follows roughly a power law with the number of distinct training environments and objects; diversity of environments and objects matters far more than simply adding more demonstrations, and returns drop off sharply once a given environment or object has enough demonstrations. Based on this, they recommend an efficient collection strategy: change environments often, pair each environment with a different object, and collect about 50 demonstrations per environment.

ExampleFollowing this recipe, data collected by 4 people in a single afternoon was enough for policies on two new tasks to reach about a 90% success rate in environments and on objects they had never seen.

Related

Scaling Law · Data Diversity · Imitation Learning · Universal Manipulation Interface · Generalization · Diffusion Policy

Sources

Data Scaling Laws in Imitation Learning for Robotic Manipulation (arXiv 2410.18647)
Data Scaling Laws in Imitation Learning 项目主页 (Chinese)

As of

2025-01